NYSS is piloting
AI marking
its English teachers check first.
A Hong Kong secondary school's English panel, and a workspace for writing, reading, listening and speaking where every AI mark arrives with the words that justify it — and nothing reaches a student until a teacher releases it.
This is a pilot, told as one. No results yet — they will be measured first.
- School
- NYSS — a Hong Kong secondary school
- Who it is for
- The English panel — its teachers and their students
- Skills
- Writing · Reading · Listening · Speaking
- Stage
- Pilot — results not yet measured
One school,
one English panel,
four skills.
- School
- NYSS, a Hong Kong secondary school
- Department
- English panel
- What we built
- An English teaching workspace: AI marking and practice that a teacher checks before any student sees it
- Who uses it
- Teachers set work, check the AI and release; students sign in with a personal code their teacher gives them
- Rubric
- HKDSE-style Content, Language and Organisation by default — for formative feedback, not official HKEAA bands — or the school’s own rubric, entered by a teacher
- Stage
- Pilot. No result from NYSS teachers has been measured yet
Detailed feedback takes
hours of marking
a teacher does not have.
Good feedback on writing means corrections in the student's own sentences and a reason for every score. Good marking of reading means checking every answer against the passage. Done by hand for a whole class, it takes hours.
The aim of the pilot: students get detailed feedback the same day, and teachers spend their time checking judgements, not writing every correction by hand — with AI marking only if a teacher can check it.
AI that proposes.
A teacher who decides.
Teachers create a class, hand out a private sign-in code to each student, and set work in any of the four skills. Within seconds of a student handing in, the AI has marked it — and the teacher sees the evidence before anyone else does.
Writing, marked with evidence
Content, Language and Organisation, each score backed by quotes from the essay and corrections shown in the student’s own text. Essays are typed or photographed — the transcription keeps the student’s own mistakes, and the student checks it before submitting. One revision included.
Real students’ handwriting is not yet tested.
Reading, marked line by line
The teacher pastes a passage and writes each question with its marking point. Every answer is marked right or wrong, with the passage line that decides it. The marking point never reaches a student before release.
Listening from a real recording
The teacher uploads a recording; the AI drafts the transcript and the teacher approves it. Students answer by ear, and the transcript stays hidden until release.
Speaking practice
Students record an answer in the browser; the AI writes down what they said and suggests marks for communication, language and ideas. The teacher listens and releases.
Pronunciation of real students is not yet measured, so the AI’s delivery suggestion goes to the teacher only.
Practice from their own mistakes
Exercises built only from each student’s verified corrections, marked by the server, not the AI. Planning help before the first draft gives questions, ideas and phrases — it never writes the essay.
The teacher releases every mark
Nothing — no mark, feedback or answer key — reaches a student until the teacher releases it. The teacher can change any mark and add a comment; every change is kept in the history.
Every quote the AI relies on is matched against the essay, the passage or the transcript. One that is not really there is withheld and the work is flagged for the teacher — never shown to the student. Totals are added up by our code, never copied from the AI.
Measured against
the teachers’ own marks.
Agreement first. Before any student uses it, the AI marks essays the English panel has already marked, and we report how closely the two agree.
Every override becomes a test case. Every change a teacher makes to an AI mark is already recorded. In the pilot those changes become test cases — so the AI is measured against the school's own standard, and each change to it is kept only if accuracy rises and nothing falls.
Until that is measured, we do not claim it.
- 100%Reading answers matched the marking key
- 38 answers × 3 runs — paraphrased, misspelled, near-miss, blank
- 7 / 7Planted grammar errors found
- every run, each with the right correction
- 0Corrections on an error-free letter
- the AI leaves correct writing alone
- ±1Same essay, marked again
- total out of 21, three runs per essay
The product's own marking code, run on 4 Oct 2026 against cases whose answers were written before the AI saw them. HKDSE-style rubric for formative feedback — not official HKEAA bands.
Pilot in progress.
Results will be published here when measured.
There are no outcomes from NYSS teachers on this page because none have been measured yet. What is not yet measured:
- 01
How closely the AI agrees with NYSS teachers’ own marks
- 02
Pronunciation of real Hong Kong students
- 03
Real students’ handwriting, photographed
- 04
Safari, iPad and phones — tested so far in Chrome on a laptop
Read the evidence.
Then talk to us.
What we have measured, with its limits, and the real screens a teacher and a student use — step by step.